Masaki Hirokawa here from Gemini Lab.
Week 2 of May is in the books. Last week's "from working to earning" arc carried straight through, and on top of it, three big waves landed at once this week: Gemini 3.2 going live, cost compression becoming a required design subject, and local LLMs finally becoming a realistic option.
When I lay this week's posts out side by side, the contours come into clear focus. Gemini 3.2 release coverage (complete guide, 3.1 / 2.5 Pro comparisons, the 7-day migration playbook), cost decisions made at design time (Caching, Cloudflare D1, cold-start mitigation, count_tokens drift), building Gemma 4 into production with OpenCode and Android Studio, and production patterns for shipping Gemini in iOS / Electron / Expo apps. With Google I/O 2026 only two weeks out, the writing felt sharper this week — and necessarily so.
Pillar 1: Gemini 3.2 Lands — A Complete Guide and a Realistic 7-Day Migration
The biggest event of the week was Gemini 3.2 going live. From the post that disambiguates the immediate questions ("which is the correct model ID?", "is the API actually compatible with 3.1?") to the implementation playbook that ships the migration in a week without dropping production, I wrote them as one connected set.
Gemini 3.2 Complete Guide — What Changed, How It Actually Feels, and How It Differs from Earlier Models captures the shifts in 3.2 with the hands-on feel layered on top of the benchmarks — the kind of "this is what actually moves the needle in real work" you don't get from numbers alone.
Built on top of that, Where Gemini 3.2 Excels and Where It Doesn't — Lessons from Real Production Use is candid about the areas where 3.2 is still weaker than expected. The honest takeaway is not "switch everything to 3.2" — there are domains where Pro / Flash / earlier models should keep their roles.
On the API side, Gemini 3.2 API Implementation Guide — Correct Model IDs, Migrating from 3.1, and Production Checkpoints and A 7-Day Production Migration Playbook from Gemini 2.5 Pro to Gemini 3.2 Pro are the two halves of the same job. The latter goes deep — prompt-compatibility checks, automated diff evaluation of outputs, and rollback design — so production traffic doesn't get dropped during the cutover.
If you want to compare side-by-side, Gemini 3.2 vs Claude Sonnet 4.6 vs GPT-4o — A Solo Developer's Honest Comparison, May 2026 is the post for you. From the perspective of three months on 3.1 Pro, Three Months on Gemini 3.1 Pro — A Solo Developer's Honest Review runs alongside it. I wrote both the parts that exceeded my expectations and the parts where, frankly, I felt let down — so it should be useful as input for buy/upgrade decisions.
Pillar 2: Cost Compression Became a Required Design Subject
The clearest shift I felt this week is that cost compression is no longer a "nice optimization" — it's a required subject you decide at design time.
Front and center is Slicing My Gemini API Bill from ¥30K/mo to ¥6K/mo with Caching: A Production Design Guide for Context Caching and Implicit Caching. The piece is the actual implementation walk-through — code plus design rationale — for compressing my own production app's monthly Gemini bill from ¥30K down to ¥6K. Implicit Caching's behavior diverges from intuition often enough that I broke its failure modes out separately into When Gemini API's Implicit Caching Doesn't Hit, or the Bill Looks Wrong: A Cause-by-Cause Troubleshooting Guide.
If you want to lean further toward the edge, Gemini API × Cloudflare D1: An Edge-SQL Production Masterclass to Keep Your AI Backend Under $10/mo and Building a Fully Edge RAG with Gemini API × Cloudflare Vectorize should both land. With Workers + D1 + Vectorize, you get "global delivery, low latency, low cost" without depending on a single region.
The cold-start problem that quietly chases away paying users is covered in Solving Gemini API's 6-Second Cold Start — Startup-Optimization Designs for Cloud Run, Lambda, and Workers.
If you want to bake free-tier mechanics and pricing into the design itself, read Understanding the Gemini API Pricing Tiers — A Cost Strategy That Squeezes the Free Tier alongside The Gemini API Prepaid Billing Migration in Full, 2026 — Service Impact and Response Checklist. The prepaid migration is one of those things that pays off if you read it once before the bill arrives and panic sets in.
Pillar 3: Gemma 4 Made Local LLMs a Realistic Option
The other strong shift this week was that Gemma 4 reached a usable bar to run locally as part of a real workflow.
Building a 'Free Claude Code' Environment with Gemma 4 26B A4B × OpenCode — Mac / Linux Hardware Setup and A Serious Use Guide for Gemma 4 × OpenCode: Building a Practical Local-LLM Dev Stack both confirm — on real hardware — that the "ask the model to write code, then iterate on it" loop is now within striking distance of Claude Code on a single laptop. The headline implementation is still the API, but being able to design "offload tasks where the cost is intolerable to the local box" is a genuinely new option.
Closer to native tooling, Running Gemma 4 in Android Studio's Local-LLM Surface — Ollama Wiring and the Real Dev Experience lines up with Gemma 4 + Ollama + Android Studio — A Complete Guide to a Local-AI Dev Environment for Real Projects. The week local LLMs moved from "experiment" to "daily driver" is how I'd want this stretch to be remembered.
For Japanese multimodal, Gemma 4 and Nemotron 3 Nano Omni — Production Design Guidelines for Japanese Multimodal AI is the one to read.
Pillar 4: Production Patterns for iOS, Mobile, and Desktop Integration
Beyond just calling the API, this week I went deep into the design of actually shipping Gemini inside an app and selling it.
The flagship for iOS is Embedding Gemini in iOS Apps with Firebase AI Logic × SwiftUI — From Architecture to App Store Release. It walks through the architecture for putting Gemini into a SwiftUI app in a form that passes App Store review, plus the pre-release checklist. Read it together with Don't Embed the Gemini API Key in Your Mobile App: A Complete Guide to Defense-in-Depth with Firebase App Check — not bundling the API key into the app is one of the issues you absolutely have to close out before charging real money.
For finer-grained iOS work, Putting the Gemini API Inside an iOS Widget: TimelineProvider Execution Limits and an App Groups Caching Pattern is this week's piece. Without respecting the TimelineProvider execution limits, widgets just quietly stop refreshing.
On cross-platform, A Complete Implementation Master Guide for Gemini API × Electron.js — Building Mac / Windows Desktop AI Apps Solo and Embedding the Gemini Live API in an Expo App — A Real-Time Voice Conversation Implementation Guide sit side by side. Both go through review, distribution, and update flows for desktop (Electron) and mobile (Expo), so they should function as input material for taking a hobby project to a real product.
Pillar 5: A Week That Thickened the Troubleshooting Layer
I also wrote a lot this week for the readers who arrived from search and need to unstick right now. The first five things to check when DEADLINE_EXCEEDED keeps firing; cause-by-cause diagnosis for FAILED_PRECONDITION; what RECITATION actually is and how to route around it; root-cause fixes for "This model is overloaded right now"; what to do when Gemini Deep Research stalls partway through; and what to check when "Gemini's answers feel worse than they used to". Each one is small on its own, but in aggregate the trust that "if I'm stuck on Gemini, gemilab.net solves it" is exactly what these unglamorous posts compound.
On the design-judgment axis, When the Gemini API's Output Just Feels Off — Seven Common Causes and a Fast-Reach Checklist is the one I expect the most readers to lean on this week.
Two Weeks Out from Google I/O 2026
This week was also pre-I/O groundwork. Two Weeks Out from Google I/O 2026 — What I'm Personally Watching for in Gemini and Two Weeks Until Google I/O 2026 — What Production Gemini API Developers Should Be Locking Down Right Now are the two pieces. Together they sketch what's likely to be announced and what you should be freezing in production today.
For the next two weeks, the writing priority is the live reaction on I/O day plus a "production-impact / no-impact" sort of every announcement.
Looking to Next Week
The week after Japan's Golden Week ends, I plan to lean into "Gemini 3.2: one month of production usage" posts and "Gemini Gems and NotebookLM — how to split work between them in real operations."
Riding into I/O 2026, the second half of May at Gemini Lab is about upgrading this week's three pillars — 3.2 landing, cost, local LLMs — into designs you can run for the long haul.
Thanks for reading this week. If something here caught your attention, please leave the tab open and come back to it when you get stuck — what keeps Gemini Lab going is the small accumulating sense that "oh, that page had me covered when something broke at 2am."
See you next week.